۱- Introduction

۲- System overview

۳- Methodology

۴- Experiments

۵- Transfer learning

۶- Conclusion

References

Abstract

Machine vision technologies have been widely used for automating the product quality control, but the defect inspection for codes on complex backgrounds is still a challenging task in the plastic container industry. In this work, an efficient and accurate inspection solution based on deep learning was proposed aiming at the detection of codes on complex backgrounds for the plastic container such as beverage packages. Firstly, image processing algorithms such as the region translation method, morphological processing, and image matching technology based on SIFT (Scale Invariant Feature Transform) features were implemented to generate synthetic defective samples, which moderated the class-imbalance problem. Data augmentation strategies were used to increase the amount of training data. Secondly, the ShuffleNet V2 framework was adapted to inspect inkjet codes on complex backgrounds. Additionally, the transfer learning was used to transfer the trained model to other inspection tasks for different kinds of packages. Finally, the proposed approach was built onto an in-line code inspection apparatus for the plastic container industry, and an accuracy of 0.9988 was achieved. The in-line testing results of false detection and omission detection rates demonstrated that the proposed solution can fully meet the production requirements. To the best of our knowledge, this report describes the first time that deep learning has been applied to the industrial defect inspection for the plastic container industry.

Introduction

The plastic packaging technologies have been experiencing explosive growth in the beverage industry. In order to ensure the product traceability and quality control, batch number and data codes are marked on the curved plastic surfaces early in the production process. The inkjet printing process, as the most commonly used type of printer, has been widely adopted to spray a digital image of the expiration date and manufacturing date by propelling droplets of ink onto the container. The quality of characters is affected by the performance of code printer and other external factors, and Legible codes on the containers provide consumers with important information and confidence for the inside product. If containers or packages with defective codes such as missing, incorrect, or unreadable codes are not identified in time, the product quality and corporate reputation will be compromised. Therefore, the code inspection that verifies the presence, position, and formation of printed codes is a key checkpoint in the plastic container industry to ensure products meet specifications prior to release and shipment to customers. With the development of machine learning and image processing technologies [1], the Optical Character Recognition (OCR) technology has been widely used by companies in the field of automatic code inspection on product containers, which greatly improves the accuracy and efficiency of code detection, and reduces their production costs as well as increases their profits. The general code detection is a composite process that comprises several phases such as the preprocessing, segmentation, feature extraction, and classification. The principal purpose is to compare the recognized characters with the correct codes and judge whether they are qualified. The widely used pixel binarization method for the character segmentation is the threshold-based segmentation method, and characters are segmented by setting a fixed or dynamic threshold.